A Digital Architecture for the Real-Time Tracking of Wearing off Phenomenon in Parkinson's Disease

Abstract

Levodopa administration is currently the most common treatment to alleviate Parkinson's Disease (PD) symptoms. Nevertheless, prolonged use of Levodopa leads to a wearing-off (WO) phenomenon, causing symptoms to reappear. To build a personalized treatment plan aiming to manage PD and its symptoms effectively, there is a need for a technological system able to continuously and objectively assess the WO phenomenon during daily life. In this context, this paper proposes a WO tracker able to exploit neuromuscular data acquired by a dedicated wireless sensor network to discriminate between a Levodopa benefit phase and the reappearance of symptoms. The proposed architecture has been implemented on a heterogeneous computing platform, that statistically analyzes neural and muscular features to identify the best set of features to train the classifier model. Eight models among shallow and deep learning approaches are analyzed in terms of performance, timing and complexity metrics to identify the best inference engine. Experimental results on five subjects experiencing WO, showed that, in the best case, the proposed WO tracker can achieve an accuracy of ~84%, providing the inference in less than 41 ms. It is possible by employing a simple fully-connected neural network with 1 hidden layer and 32 units.

Keywords: EEG; EMG; Parkinson’s Disease; heterogeneous computing platform; wearing off.

Figure 1

Overall architecture of the Wearing Off (WO) detector, considering an implementation of the heterogeneous computing platform by [6]. The here-proposed approach implements on a Field Programmable Gate Array (FPGA) the feature extraction phase concerning the diagnostic indexes from muscles and brain, via Movement Related Potentials (MRP) as in [6]. The classification phase is instead entrusted to the device microcontroller (μC).

Figure 2

Bar chart for selected muscular features. Bar values identify the mean value for the considered parameter. Error bar limits represent the 95th percentile (upper bound) and 5th percentile (lower bound). LG_TA is the co-contraction time (ms) between Lateral Gatrocnemius (LG) and Tibialis Anterior (TA), RF_BF identifies the co-contraction time (ms) between Rectus Femoris (RF) and Bicep Femoris (BF) regardless the involved side.

Figure 3

Boxplot representation of the Testing (blue boxplot) and Training (red boxplot) Set features distribution. The middle line of the boxplot represents the median value of the distribution, upper and lower boundaries of the boxplot are respectively the 75^th and the 25^th percentile of the considered sample. The upper and lower limits of the bar are the maximum and minimum adjacent, while the circles denote outliers. LG_TA identifies the co-contraction time (ms) between Lateral Gatrocnemius (LG) and Tibialis Anterior (TA), similarly, RF_BF concerns Rectus Femoris (RF) and Bicep Femoris (BF) muscles regardless of the involved side.

Figure 4

Shallow Learning Classification model performances: Confusion matrix, Receiver Operating Characteristic (ROC) curves and Area Under Curve (AUC) parameter. The positive class selected for the reported computation is the OFF phenomenon of Wearing Off (WO). The red dot on the ROC denotes the selected and analyzed classifier, minimizing the distance among the True Positive Rate (TPR) and False Positive Rate (FPR) values of the ROC curve and {FPR, TPR} = {0,1}.

Figure 5

Deep Learning Classification model performances: Confusion matrix, Receiver Operating Characteristic (ROC) curves and Area Under Curve (AUC) parameter. The positive class selected for the reported computation is the OFF phenomenon of Wearing Off (WO). The red dot on the ROC denotes the selected and analyzed classifier, minimizing the distance among the True Positive Rate (TPR) and False Positive Rate (FPR) values of the ROC curve and {FPR, TPR} = {0,1}.